LLM Summary Generation With Hallucination Filtering and Section Prompts
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Large language models face challenges in processing complex texts like clinical study reports (CSRs) due to extreme latency, computational expense, and introduction of hallucinations, making it difficult to generate accurate and simplified summaries.
Innovation Solution
Implementing a series of classifications, scores, and prompts to create curated inputs for large language models, along with filters to remove hallucinations and align content sections with appropriate prompts, thereby reducing latency and power consumption while ensuring accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If large language models process large amounts of complex text directly, then comprehensive summarization is achieved, but extreme latency and computational expense occur
Solution Approach 1:
The patent segments the complex text into smaller sections and processes them individually through multiple passes. Each pass focuses on specific sections, breaking down the large text processing task into manageable chunks that can be handled more efficiently by the language model, thereby reducing overall latency while maintaining comprehensive coverage.
Solution Approach 2:
The system performs preliminary text processing and analysis before feeding content to the large language model. This includes initial parsing, section identification, and preparation of structured inputs that reduce the complexity of the main summarization task, allowing the model to work more efficiently on pre-processed data.
2Measurement precision
If large language models process large amounts of complex text directly, then comprehensive summarization is achieved, but computational expense increases
Solution Approach 1:
By dividing the text processing into segments and passes, the system reduces the computational burden on the large language model at any given time. Each pass processes a subset of the content, requiring fewer computational resources per invocation while achieving comprehensive coverage through multiple targeted passes.
Solution Approach 2:
The system performs multiple passes over different sections of the text rather than attempting to process everything in a single comprehensive pass. This partial action approach allows the model to focus computational resources on specific sections that require attention, reducing overall computational expense while maintaining accuracy.
3Productivity
If large language models generate summaries from complex text, then summarization is produced, but hallucinations are introduced
Solution Approach 1:
The patent implements a feedback mechanism where the generated summary is validated against the original text through classifier models. These classifiers check whether summary statements are supported by evidence in the source text, providing feedback that identifies and flags potential hallucinations for correction or removal.
Solution Approach 2:
The system performs preliminary validation and fact-checking of summary content using classifier models before finalizing the output. This preliminary action of verifying claims against the source text prevents hallucinations from appearing in the final summary, maintaining reliability while preserving the efficiency of automated generation.
4Measurement precision
If complex texts are summarized maintaining original reading level, then technical accuracy is preserved, but accessibility to broader audience is reduced
Solution Approach 1:
The patent applies different processing qualities to different parts of the summarization task. Technical terms and concepts are preserved with their original meaning and accuracy, while the surrounding explanatory text is simplified in language and structure. This local differentiation allows the summary to maintain technical accuracy where needed while improving accessibility elsewhere.
Solution Approach 2:
The system changes the language complexity parameter of the summary output while preserving the core technical information. By adjusting vocabulary difficulty, sentence structure complexity, and explanatory depth, the system produces summaries that are accessible to broader audiences without sacrificing the accuracy of technical content through careful parameter control.
Data Source
AI summary
Systems and methods for generating summaries from text using a generative model are disclosed. The system is configured to access an article; identify section; provide, to one or more generative models, a prompt including instructions to generate a section summary; generate an article summary based on the section summary; determine, from the article summary, a first concept found in the article summary that is missing from the article; determine, using a classifier, for a first sentence included in the article summary, a confidence score; and provide, for presentation at a client device, a document including the article summary.